AI Video Dubber & Translator

Takes a video in one language and gives you back the same video speaking another. It transcribes the original speech, translates the text, generates a fresh AI voiceover timed to the original pacing, and muxes that audio back onto the video. Built for creators, localization teams, and anyone who needs a quick second-language cut without booking voice talent.

There is no voice cloning here. The output uses a standard synthetic voice, so you don't need anyone's consent to ship it.

$0.12 per dubbed minute, plus $0.0000125 each time a run starts. Flat rate — no volume tiers, no plan gates — and you are charged only for minutes actually dubbed.

How it works

Four stages, in order:

1. Transcribe — the source audio is pulled out and sent to Whisper, which returns time-coded segments. Set sourceLanguage to auto and it detects the spoken language for you.
2. Translate — each segment is translated by a chat model (gpt-4o-mini by default, overridable via translationModel), segment by segment so the timings survive.
3. Speak — OpenAI TTS renders the translated text in the voice you picked.
4. Assemble — the generated clips are fitted back to the original segment timings, then muxed onto the source video with ffmpeg. If burnSubtitles is on, the translated lines are rendered into the frame as styled ASS subtitles rather than added as a soft track.

Input

Field	Required	Notes
videoUrl	yes	Public direct URL to the source video (.mp4, .mov, .webm). You host it; nothing is scraped.
openaiApiKey	yes	Your own OpenAI key. Used for transcription, translation and TTS. Marked secret in the schema.
targetLanguage	no	23 options: Spanish, French, German, Italian, Portuguese, Hindi, Arabic, Russian, Japanese, Korean, Chinese, Dutch, Polish, Turkish, Indonesian, Vietnamese, Thai, Swedish, Ukrainian, Romanian, Greek, Hebrew, Filipino. Defaults to es.
sourceLanguage	no	ISO code, or auto to detect. Default auto.
voice	no	alloy, echo, fable, onyx (deep male, default), nova (female), shimmer (female).
burnSubtitles	no	Overlay the translated subtitles onto the video. On by default.
ttsModel	no	tts-1 (fast, default) or tts-1-hd (higher quality).
translationModel	no	Chat model used for translation. Default gpt-4o-mini.
baseUrl	no	Any OpenAI-compatible base URL, if you route through a proxy or a compatible provider.

Output

One dataset row per run, plus the media in the run's key-value store.

Field	What it is
ok	Whether the dub completed.
videoUrl	The source you passed in.
sourceLanguage	The language actually detected (or the one you forced).
targetLanguage, voice	What was produced, echoed back.
segments	How many timed speech segments were transcribed, translated and re-voiced.
durationSeconds	Source duration, rounded to the second. This is what billing is derived from.
output.mp4Key	Key-value-store key of the finished video, e.g. dubbed-es-1754689200000.mp4.
output.srtKey	Key of the translated .srt, written on every run whether or not subtitles were burned in.
output.mp4Url	Direct URL to the finished MP4 when the store exposes one.
processingSeconds	Wall-clock time the run took.

Example

{
  "videoUrl": "https://example.com/clip.mp4",
  "targetLanguage": "es",
  "sourceLanguage": "auto",
  "voice": "onyx",
  "burnSubtitles": true,
  "ttsModel": "tts-1",
  "openaiApiKey": "sk-..."
}

Pricing

Event	When	Price
Actor start	Once per run, per GB of run memory	$0.0000125
Dubbed minute	Per minute of dubbed video	$0.12

Minutes are rounded up to the next whole minute, with a one-minute minimum — a 90-second clip is charged as two minutes, $0.24. The dub-minute event fires only after the finished MP4 has been written, so a run that fails partway through costs you the start fee and nothing else.

Run it without a videoUrl or without a key and it returns a single row labelled _sample: true showing the output shape. Sample runs are not charged for dubbed minutes.

Your OpenAI usage is billed separately by OpenAI, on your own key. That covers Whisper, the translation model and TTS.

Notes and limits

• The voiceover is fitted to the original segment timings. If the target language is much wordier than the source — English into German is the usual offender — pacing on long segments can feel rushed.
• Burned subtitles are rendered into the pixels and cannot be toggled off in a player. Turn burnSubtitles off and use the .srt from the key-value store if you want a soft track.
• One video per run. Queue several runs to batch a catalog.
• The source must be a direct, publicly reachable file URL. Page URLs, playlists and login-gated links will not download.